Semi-Supervised Multi-View Ensemble Clustering with Subspaces Extraction and Constraints Propagation

Semi-supervised multi-view clustering extends the principles of multi-modal analysis by focusing on the partitioning of data into distinct groups based on limited number of pairwise constraints. Guided by pairwise constraints, the existing methods suffer from a lack of integration at the level of view-specific subspaces and partitions, which may lead to information loss of each view. In order to sovle this problem, this paper proposes a semi-supervised multi-view ensemble clustering framework with subspaces extraction and constraints propagation (SSMVEC). For each view, a constrained multi-view broad self-encoding network (CMVBSN) is firstly designed to integrate cross-view affinity, within-view projection and pairwise constraints to extract a compressed feature subspace and derive a base partition. By absorbing the information from both pairwise constraints and multiple partitions, the bidirectional constraint propagation is adopted to refines the unified similarity matrix. Finally, the graph-cut algorithm is applied to obtain the consensus partition. Demonstrated by the analysis of parameters, comparison methods and incremental constraints on multiple multi-view datasets, SSMVEC can effectively leverage pairwise constraints and achieve overall better performance than several multi-view clustering methods.